Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input
[linux-2.6.git] / mm / vmscan.c
1 /*
2  *  linux/mm/vmscan.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *
6  *  Swap reorganised 29.12.95, Stephen Tweedie.
7  *  kswapd added: 7.1.96  sct
8  *  Removed kswapd_ctl limits, and swap out as many pages as needed
9  *  to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10  *  Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11  *  Multiqueue VM started 5.8.00, Rik van Riel.
12  */
13
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/slab.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h>  /* for try_to_release_page(),
27                                         buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/pagevec.h>
30 #include <linux/backing-dev.h>
31 #include <linux/rmap.h>
32 #include <linux/topology.h>
33 #include <linux/cpu.h>
34 #include <linux/cpuset.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40
41 #include <asm/tlbflush.h>
42 #include <asm/div64.h>
43
44 #include <linux/swapops.h>
45
46 #include "internal.h"
47
48 struct scan_control {
49         /* Incremented by the number of inactive pages that were scanned */
50         unsigned long nr_scanned;
51
52         /* This context's GFP mask */
53         gfp_t gfp_mask;
54
55         int may_writepage;
56
57         /* Can pages be swapped as part of reclaim? */
58         int may_swap;
59
60         /* This context's SWAP_CLUSTER_MAX. If freeing memory for
61          * suspend, we effectively ignore SWAP_CLUSTER_MAX.
62          * In this context, it doesn't matter that we scan the
63          * whole list at once. */
64         int swap_cluster_max;
65
66         int swappiness;
67
68         int all_unreclaimable;
69
70         int order;
71 };
72
73 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
74
75 #ifdef ARCH_HAS_PREFETCH
76 #define prefetch_prev_lru_page(_page, _base, _field)                    \
77         do {                                                            \
78                 if ((_page)->lru.prev != _base) {                       \
79                         struct page *prev;                              \
80                                                                         \
81                         prev = lru_to_page(&(_page->lru));              \
82                         prefetch(&prev->_field);                        \
83                 }                                                       \
84         } while (0)
85 #else
86 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
87 #endif
88
89 #ifdef ARCH_HAS_PREFETCHW
90 #define prefetchw_prev_lru_page(_page, _base, _field)                   \
91         do {                                                            \
92                 if ((_page)->lru.prev != _base) {                       \
93                         struct page *prev;                              \
94                                                                         \
95                         prev = lru_to_page(&(_page->lru));              \
96                         prefetchw(&prev->_field);                       \
97                 }                                                       \
98         } while (0)
99 #else
100 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
101 #endif
102
103 /*
104  * From 0 .. 100.  Higher means more swappy.
105  */
106 int vm_swappiness = 60;
107 long vm_total_pages;    /* The total number of pages which the VM controls */
108
109 static LIST_HEAD(shrinker_list);
110 static DECLARE_RWSEM(shrinker_rwsem);
111
112 /*
113  * Add a shrinker callback to be called from the vm
114  */
115 void register_shrinker(struct shrinker *shrinker)
116 {
117         shrinker->nr = 0;
118         down_write(&shrinker_rwsem);
119         list_add_tail(&shrinker->list, &shrinker_list);
120         up_write(&shrinker_rwsem);
121 }
122 EXPORT_SYMBOL(register_shrinker);
123
124 /*
125  * Remove one
126  */
127 void unregister_shrinker(struct shrinker *shrinker)
128 {
129         down_write(&shrinker_rwsem);
130         list_del(&shrinker->list);
131         up_write(&shrinker_rwsem);
132 }
133 EXPORT_SYMBOL(unregister_shrinker);
134
135 #define SHRINK_BATCH 128
136 /*
137  * Call the shrink functions to age shrinkable caches
138  *
139  * Here we assume it costs one seek to replace a lru page and that it also
140  * takes a seek to recreate a cache object.  With this in mind we age equal
141  * percentages of the lru and ageable caches.  This should balance the seeks
142  * generated by these structures.
143  *
144  * If the vm encountered mapped pages on the LRU it increase the pressure on
145  * slab to avoid swapping.
146  *
147  * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
148  *
149  * `lru_pages' represents the number of on-LRU pages in all the zones which
150  * are eligible for the caller's allocation attempt.  It is used for balancing
151  * slab reclaim versus page reclaim.
152  *
153  * Returns the number of slab objects which we shrunk.
154  */
155 unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
156                         unsigned long lru_pages)
157 {
158         struct shrinker *shrinker;
159         unsigned long ret = 0;
160
161         if (scanned == 0)
162                 scanned = SWAP_CLUSTER_MAX;
163
164         if (!down_read_trylock(&shrinker_rwsem))
165                 return 1;       /* Assume we'll be able to shrink next time */
166
167         list_for_each_entry(shrinker, &shrinker_list, list) {
168                 unsigned long long delta;
169                 unsigned long total_scan;
170                 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
171
172                 delta = (4 * scanned) / shrinker->seeks;
173                 delta *= max_pass;
174                 do_div(delta, lru_pages + 1);
175                 shrinker->nr += delta;
176                 if (shrinker->nr < 0) {
177                         printk(KERN_ERR "%s: nr=%ld\n",
178                                         __FUNCTION__, shrinker->nr);
179                         shrinker->nr = max_pass;
180                 }
181
182                 /*
183                  * Avoid risking looping forever due to too large nr value:
184                  * never try to free more than twice the estimate number of
185                  * freeable entries.
186                  */
187                 if (shrinker->nr > max_pass * 2)
188                         shrinker->nr = max_pass * 2;
189
190                 total_scan = shrinker->nr;
191                 shrinker->nr = 0;
192
193                 while (total_scan >= SHRINK_BATCH) {
194                         long this_scan = SHRINK_BATCH;
195                         int shrink_ret;
196                         int nr_before;
197
198                         nr_before = (*shrinker->shrink)(0, gfp_mask);
199                         shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
200                         if (shrink_ret == -1)
201                                 break;
202                         if (shrink_ret < nr_before)
203                                 ret += nr_before - shrink_ret;
204                         count_vm_events(SLABS_SCANNED, this_scan);
205                         total_scan -= this_scan;
206
207                         cond_resched();
208                 }
209
210                 shrinker->nr += total_scan;
211         }
212         up_read(&shrinker_rwsem);
213         return ret;
214 }
215
216 /* Called without lock on whether page is mapped, so answer is unstable */
217 static inline int page_mapping_inuse(struct page *page)
218 {
219         struct address_space *mapping;
220
221         /* Page is in somebody's page tables. */
222         if (page_mapped(page))
223                 return 1;
224
225         /* Be more reluctant to reclaim swapcache than pagecache */
226         if (PageSwapCache(page))
227                 return 1;
228
229         mapping = page_mapping(page);
230         if (!mapping)
231                 return 0;
232
233         /* File is mmap'd by somebody? */
234         return mapping_mapped(mapping);
235 }
236
237 static inline int is_page_cache_freeable(struct page *page)
238 {
239         return page_count(page) - !!PagePrivate(page) == 2;
240 }
241
242 static int may_write_to_queue(struct backing_dev_info *bdi)
243 {
244         if (current->flags & PF_SWAPWRITE)
245                 return 1;
246         if (!bdi_write_congested(bdi))
247                 return 1;
248         if (bdi == current->backing_dev_info)
249                 return 1;
250         return 0;
251 }
252
253 /*
254  * We detected a synchronous write error writing a page out.  Probably
255  * -ENOSPC.  We need to propagate that into the address_space for a subsequent
256  * fsync(), msync() or close().
257  *
258  * The tricky part is that after writepage we cannot touch the mapping: nothing
259  * prevents it from being freed up.  But we have a ref on the page and once
260  * that page is locked, the mapping is pinned.
261  *
262  * We're allowed to run sleeping lock_page() here because we know the caller has
263  * __GFP_FS.
264  */
265 static void handle_write_error(struct address_space *mapping,
266                                 struct page *page, int error)
267 {
268         lock_page(page);
269         if (page_mapping(page) == mapping)
270                 mapping_set_error(mapping, error);
271         unlock_page(page);
272 }
273
274 /* Request for sync pageout. */
275 enum pageout_io {
276         PAGEOUT_IO_ASYNC,
277         PAGEOUT_IO_SYNC,
278 };
279
280 /* possible outcome of pageout() */
281 typedef enum {
282         /* failed to write page out, page is locked */
283         PAGE_KEEP,
284         /* move page to the active list, page is locked */
285         PAGE_ACTIVATE,
286         /* page has been sent to the disk successfully, page is unlocked */
287         PAGE_SUCCESS,
288         /* page is clean and locked */
289         PAGE_CLEAN,
290 } pageout_t;
291
292 /*
293  * pageout is called by shrink_page_list() for each dirty page.
294  * Calls ->writepage().
295  */
296 static pageout_t pageout(struct page *page, struct address_space *mapping,
297                                                 enum pageout_io sync_writeback)
298 {
299         /*
300          * If the page is dirty, only perform writeback if that write
301          * will be non-blocking.  To prevent this allocation from being
302          * stalled by pagecache activity.  But note that there may be
303          * stalls if we need to run get_block().  We could test
304          * PagePrivate for that.
305          *
306          * If this process is currently in generic_file_write() against
307          * this page's queue, we can perform writeback even if that
308          * will block.
309          *
310          * If the page is swapcache, write it back even if that would
311          * block, for some throttling. This happens by accident, because
312          * swap_backing_dev_info is bust: it doesn't reflect the
313          * congestion state of the swapdevs.  Easy to fix, if needed.
314          * See swapfile.c:page_queue_congested().
315          */
316         if (!is_page_cache_freeable(page))
317                 return PAGE_KEEP;
318         if (!mapping) {
319                 /*
320                  * Some data journaling orphaned pages can have
321                  * page->mapping == NULL while being dirty with clean buffers.
322                  */
323                 if (PagePrivate(page)) {
324                         if (try_to_free_buffers(page)) {
325                                 ClearPageDirty(page);
326                                 printk("%s: orphaned page\n", __FUNCTION__);
327                                 return PAGE_CLEAN;
328                         }
329                 }
330                 return PAGE_KEEP;
331         }
332         if (mapping->a_ops->writepage == NULL)
333                 return PAGE_ACTIVATE;
334         if (!may_write_to_queue(mapping->backing_dev_info))
335                 return PAGE_KEEP;
336
337         if (clear_page_dirty_for_io(page)) {
338                 int res;
339                 struct writeback_control wbc = {
340                         .sync_mode = WB_SYNC_NONE,
341                         .nr_to_write = SWAP_CLUSTER_MAX,
342                         .range_start = 0,
343                         .range_end = LLONG_MAX,
344                         .nonblocking = 1,
345                         .for_reclaim = 1,
346                 };
347
348                 SetPageReclaim(page);
349                 res = mapping->a_ops->writepage(page, &wbc);
350                 if (res < 0)
351                         handle_write_error(mapping, page, res);
352                 if (res == AOP_WRITEPAGE_ACTIVATE) {
353                         ClearPageReclaim(page);
354                         return PAGE_ACTIVATE;
355                 }
356
357                 /*
358                  * Wait on writeback if requested to. This happens when
359                  * direct reclaiming a large contiguous area and the
360                  * first attempt to free a range of pages fails.
361                  */
362                 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
363                         wait_on_page_writeback(page);
364
365                 if (!PageWriteback(page)) {
366                         /* synchronous write or broken a_ops? */
367                         ClearPageReclaim(page);
368                 }
369                 inc_zone_page_state(page, NR_VMSCAN_WRITE);
370                 return PAGE_SUCCESS;
371         }
372
373         return PAGE_CLEAN;
374 }
375
376 /*
377  * Attempt to detach a locked page from its ->mapping.  If it is dirty or if
378  * someone else has a ref on the page, abort and return 0.  If it was
379  * successfully detached, return 1.  Assumes the caller has a single ref on
380  * this page.
381  */
382 int remove_mapping(struct address_space *mapping, struct page *page)
383 {
384         BUG_ON(!PageLocked(page));
385         BUG_ON(mapping != page_mapping(page));
386
387         write_lock_irq(&mapping->tree_lock);
388         /*
389          * The non racy check for a busy page.
390          *
391          * Must be careful with the order of the tests. When someone has
392          * a ref to the page, it may be possible that they dirty it then
393          * drop the reference. So if PageDirty is tested before page_count
394          * here, then the following race may occur:
395          *
396          * get_user_pages(&page);
397          * [user mapping goes away]
398          * write_to(page);
399          *                              !PageDirty(page)    [good]
400          * SetPageDirty(page);
401          * put_page(page);
402          *                              !page_count(page)   [good, discard it]
403          *
404          * [oops, our write_to data is lost]
405          *
406          * Reversing the order of the tests ensures such a situation cannot
407          * escape unnoticed. The smp_rmb is needed to ensure the page->flags
408          * load is not satisfied before that of page->_count.
409          *
410          * Note that if SetPageDirty is always performed via set_page_dirty,
411          * and thus under tree_lock, then this ordering is not required.
412          */
413         if (unlikely(page_count(page) != 2))
414                 goto cannot_free;
415         smp_rmb();
416         if (unlikely(PageDirty(page)))
417                 goto cannot_free;
418
419         if (PageSwapCache(page)) {
420                 swp_entry_t swap = { .val = page_private(page) };
421                 __delete_from_swap_cache(page);
422                 write_unlock_irq(&mapping->tree_lock);
423                 swap_free(swap);
424                 __put_page(page);       /* The pagecache ref */
425                 return 1;
426         }
427
428         __remove_from_page_cache(page);
429         write_unlock_irq(&mapping->tree_lock);
430         __put_page(page);
431         return 1;
432
433 cannot_free:
434         write_unlock_irq(&mapping->tree_lock);
435         return 0;
436 }
437
438 /*
439  * shrink_page_list() returns the number of reclaimed pages
440  */
441 static unsigned long shrink_page_list(struct list_head *page_list,
442                                         struct scan_control *sc,
443                                         enum pageout_io sync_writeback)
444 {
445         LIST_HEAD(ret_pages);
446         struct pagevec freed_pvec;
447         int pgactivate = 0;
448         unsigned long nr_reclaimed = 0;
449
450         cond_resched();
451
452         pagevec_init(&freed_pvec, 1);
453         while (!list_empty(page_list)) {
454                 struct address_space *mapping;
455                 struct page *page;
456                 int may_enter_fs;
457                 int referenced;
458
459                 cond_resched();
460
461                 page = lru_to_page(page_list);
462                 list_del(&page->lru);
463
464                 if (TestSetPageLocked(page))
465                         goto keep;
466
467                 VM_BUG_ON(PageActive(page));
468
469                 sc->nr_scanned++;
470
471                 if (!sc->may_swap && page_mapped(page))
472                         goto keep_locked;
473
474                 /* Double the slab pressure for mapped and swapcache pages */
475                 if (page_mapped(page) || PageSwapCache(page))
476                         sc->nr_scanned++;
477
478                 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
479                         (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
480
481                 if (PageWriteback(page)) {
482                         /*
483                          * Synchronous reclaim is performed in two passes,
484                          * first an asynchronous pass over the list to
485                          * start parallel writeback, and a second synchronous
486                          * pass to wait for the IO to complete.  Wait here
487                          * for any page for which writeback has already
488                          * started.
489                          */
490                         if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
491                                 wait_on_page_writeback(page);
492                         else
493                                 goto keep_locked;
494                 }
495
496                 referenced = page_referenced(page, 1);
497                 /* In active use or really unfreeable?  Activate it. */
498                 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
499                                         referenced && page_mapping_inuse(page))
500                         goto activate_locked;
501
502 #ifdef CONFIG_SWAP
503                 /*
504                  * Anonymous process memory has backing store?
505                  * Try to allocate it some swap space here.
506                  */
507                 if (PageAnon(page) && !PageSwapCache(page))
508                         if (!add_to_swap(page, GFP_ATOMIC))
509                                 goto activate_locked;
510 #endif /* CONFIG_SWAP */
511
512                 mapping = page_mapping(page);
513
514                 /*
515                  * The page is mapped into the page tables of one or more
516                  * processes. Try to unmap it here.
517                  */
518                 if (page_mapped(page) && mapping) {
519                         switch (try_to_unmap(page, 0)) {
520                         case SWAP_FAIL:
521                                 goto activate_locked;
522                         case SWAP_AGAIN:
523                                 goto keep_locked;
524                         case SWAP_SUCCESS:
525                                 ; /* try to free the page below */
526                         }
527                 }
528
529                 if (PageDirty(page)) {
530                         if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
531                                 goto keep_locked;
532                         if (!may_enter_fs)
533                                 goto keep_locked;
534                         if (!sc->may_writepage)
535                                 goto keep_locked;
536
537                         /* Page is dirty, try to write it out here */
538                         switch (pageout(page, mapping, sync_writeback)) {
539                         case PAGE_KEEP:
540                                 goto keep_locked;
541                         case PAGE_ACTIVATE:
542                                 goto activate_locked;
543                         case PAGE_SUCCESS:
544                                 if (PageWriteback(page) || PageDirty(page))
545                                         goto keep;
546                                 /*
547                                  * A synchronous write - probably a ramdisk.  Go
548                                  * ahead and try to reclaim the page.
549                                  */
550                                 if (TestSetPageLocked(page))
551                                         goto keep;
552                                 if (PageDirty(page) || PageWriteback(page))
553                                         goto keep_locked;
554                                 mapping = page_mapping(page);
555                         case PAGE_CLEAN:
556                                 ; /* try to free the page below */
557                         }
558                 }
559
560                 /*
561                  * If the page has buffers, try to free the buffer mappings
562                  * associated with this page. If we succeed we try to free
563                  * the page as well.
564                  *
565                  * We do this even if the page is PageDirty().
566                  * try_to_release_page() does not perform I/O, but it is
567                  * possible for a page to have PageDirty set, but it is actually
568                  * clean (all its buffers are clean).  This happens if the
569                  * buffers were written out directly, with submit_bh(). ext3
570                  * will do this, as well as the blockdev mapping. 
571                  * try_to_release_page() will discover that cleanness and will
572                  * drop the buffers and mark the page clean - it can be freed.
573                  *
574                  * Rarely, pages can have buffers and no ->mapping.  These are
575                  * the pages which were not successfully invalidated in
576                  * truncate_complete_page().  We try to drop those buffers here
577                  * and if that worked, and the page is no longer mapped into
578                  * process address space (page_count == 1) it can be freed.
579                  * Otherwise, leave the page on the LRU so it is swappable.
580                  */
581                 if (PagePrivate(page)) {
582                         if (!try_to_release_page(page, sc->gfp_mask))
583                                 goto activate_locked;
584                         if (!mapping && page_count(page) == 1)
585                                 goto free_it;
586                 }
587
588                 if (!mapping || !remove_mapping(mapping, page))
589                         goto keep_locked;
590
591 free_it:
592                 unlock_page(page);
593                 nr_reclaimed++;
594                 if (!pagevec_add(&freed_pvec, page))
595                         __pagevec_release_nonlru(&freed_pvec);
596                 continue;
597
598 activate_locked:
599                 SetPageActive(page);
600                 pgactivate++;
601 keep_locked:
602                 unlock_page(page);
603 keep:
604                 list_add(&page->lru, &ret_pages);
605                 VM_BUG_ON(PageLRU(page));
606         }
607         list_splice(&ret_pages, page_list);
608         if (pagevec_count(&freed_pvec))
609                 __pagevec_release_nonlru(&freed_pvec);
610         count_vm_events(PGACTIVATE, pgactivate);
611         return nr_reclaimed;
612 }
613
614 /* LRU Isolation modes. */
615 #define ISOLATE_INACTIVE 0      /* Isolate inactive pages. */
616 #define ISOLATE_ACTIVE 1        /* Isolate active pages. */
617 #define ISOLATE_BOTH 2          /* Isolate both active and inactive pages. */
618
619 /*
620  * Attempt to remove the specified page from its LRU.  Only take this page
621  * if it is of the appropriate PageActive status.  Pages which are being
622  * freed elsewhere are also ignored.
623  *
624  * page:        page to consider
625  * mode:        one of the LRU isolation modes defined above
626  *
627  * returns 0 on success, -ve errno on failure.
628  */
629 static int __isolate_lru_page(struct page *page, int mode)
630 {
631         int ret = -EINVAL;
632
633         /* Only take pages on the LRU. */
634         if (!PageLRU(page))
635                 return ret;
636
637         /*
638          * When checking the active state, we need to be sure we are
639          * dealing with comparible boolean values.  Take the logical not
640          * of each.
641          */
642         if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
643                 return ret;
644
645         ret = -EBUSY;
646         if (likely(get_page_unless_zero(page))) {
647                 /*
648                  * Be careful not to clear PageLRU until after we're
649                  * sure the page is not being freed elsewhere -- the
650                  * page release code relies on it.
651                  */
652                 ClearPageLRU(page);
653                 ret = 0;
654         }
655
656         return ret;
657 }
658
659 /*
660  * zone->lru_lock is heavily contended.  Some of the functions that
661  * shrink the lists perform better by taking out a batch of pages
662  * and working on them outside the LRU lock.
663  *
664  * For pagecache intensive workloads, this function is the hottest
665  * spot in the kernel (apart from copy_*_user functions).
666  *
667  * Appropriate locks must be held before calling this function.
668  *
669  * @nr_to_scan: The number of pages to look through on the list.
670  * @src:        The LRU list to pull pages off.
671  * @dst:        The temp list to put pages on to.
672  * @scanned:    The number of pages that were scanned.
673  * @order:      The caller's attempted allocation order
674  * @mode:       One of the LRU isolation modes
675  *
676  * returns how many pages were moved onto *@dst.
677  */
678 static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
679                 struct list_head *src, struct list_head *dst,
680                 unsigned long *scanned, int order, int mode)
681 {
682         unsigned long nr_taken = 0;
683         unsigned long scan;
684
685         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
686                 struct page *page;
687                 unsigned long pfn;
688                 unsigned long end_pfn;
689                 unsigned long page_pfn;
690                 int zone_id;
691
692                 page = lru_to_page(src);
693                 prefetchw_prev_lru_page(page, src, flags);
694
695                 VM_BUG_ON(!PageLRU(page));
696
697                 switch (__isolate_lru_page(page, mode)) {
698                 case 0:
699                         list_move(&page->lru, dst);
700                         nr_taken++;
701                         break;
702
703                 case -EBUSY:
704                         /* else it is being freed elsewhere */
705                         list_move(&page->lru, src);
706                         continue;
707
708                 default:
709                         BUG();
710                 }
711
712                 if (!order)
713                         continue;
714
715                 /*
716                  * Attempt to take all pages in the order aligned region
717                  * surrounding the tag page.  Only take those pages of
718                  * the same active state as that tag page.  We may safely
719                  * round the target page pfn down to the requested order
720                  * as the mem_map is guarenteed valid out to MAX_ORDER,
721                  * where that page is in a different zone we will detect
722                  * it from its zone id and abort this block scan.
723                  */
724                 zone_id = page_zone_id(page);
725                 page_pfn = page_to_pfn(page);
726                 pfn = page_pfn & ~((1 << order) - 1);
727                 end_pfn = pfn + (1 << order);
728                 for (; pfn < end_pfn; pfn++) {
729                         struct page *cursor_page;
730
731                         /* The target page is in the block, ignore it. */
732                         if (unlikely(pfn == page_pfn))
733                                 continue;
734
735                         /* Avoid holes within the zone. */
736                         if (unlikely(!pfn_valid_within(pfn)))
737                                 break;
738
739                         cursor_page = pfn_to_page(pfn);
740                         /* Check that we have not crossed a zone boundary. */
741                         if (unlikely(page_zone_id(cursor_page) != zone_id))
742                                 continue;
743                         switch (__isolate_lru_page(cursor_page, mode)) {
744                         case 0:
745                                 list_move(&cursor_page->lru, dst);
746                                 nr_taken++;
747                                 scan++;
748                                 break;
749
750                         case -EBUSY:
751                                 /* else it is being freed elsewhere */
752                                 list_move(&cursor_page->lru, src);
753                         default:
754                                 break;
755                         }
756                 }
757         }
758
759         *scanned = scan;
760         return nr_taken;
761 }
762
763 /*
764  * clear_active_flags() is a helper for shrink_active_list(), clearing
765  * any active bits from the pages in the list.
766  */
767 static unsigned long clear_active_flags(struct list_head *page_list)
768 {
769         int nr_active = 0;
770         struct page *page;
771
772         list_for_each_entry(page, page_list, lru)
773                 if (PageActive(page)) {
774                         ClearPageActive(page);
775                         nr_active++;
776                 }
777
778         return nr_active;
779 }
780
781 /*
782  * shrink_inactive_list() is a helper for shrink_zone().  It returns the number
783  * of reclaimed pages
784  */
785 static unsigned long shrink_inactive_list(unsigned long max_scan,
786                                 struct zone *zone, struct scan_control *sc)
787 {
788         LIST_HEAD(page_list);
789         struct pagevec pvec;
790         unsigned long nr_scanned = 0;
791         unsigned long nr_reclaimed = 0;
792
793         pagevec_init(&pvec, 1);
794
795         lru_add_drain();
796         spin_lock_irq(&zone->lru_lock);
797         do {
798                 struct page *page;
799                 unsigned long nr_taken;
800                 unsigned long nr_scan;
801                 unsigned long nr_freed;
802                 unsigned long nr_active;
803
804                 nr_taken = isolate_lru_pages(sc->swap_cluster_max,
805                              &zone->inactive_list,
806                              &page_list, &nr_scan, sc->order,
807                              (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
808                                              ISOLATE_BOTH : ISOLATE_INACTIVE);
809                 nr_active = clear_active_flags(&page_list);
810                 __count_vm_events(PGDEACTIVATE, nr_active);
811
812                 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
813                 __mod_zone_page_state(zone, NR_INACTIVE,
814                                                 -(nr_taken - nr_active));
815                 zone->pages_scanned += nr_scan;
816                 spin_unlock_irq(&zone->lru_lock);
817
818                 nr_scanned += nr_scan;
819                 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
820
821                 /*
822                  * If we are direct reclaiming for contiguous pages and we do
823                  * not reclaim everything in the list, try again and wait
824                  * for IO to complete. This will stall high-order allocations
825                  * but that should be acceptable to the caller
826                  */
827                 if (nr_freed < nr_taken && !current_is_kswapd() &&
828                                         sc->order > PAGE_ALLOC_COSTLY_ORDER) {
829                         congestion_wait(WRITE, HZ/10);
830
831                         /*
832                          * The attempt at page out may have made some
833                          * of the pages active, mark them inactive again.
834                          */
835                         nr_active = clear_active_flags(&page_list);
836                         count_vm_events(PGDEACTIVATE, nr_active);
837
838                         nr_freed += shrink_page_list(&page_list, sc,
839                                                         PAGEOUT_IO_SYNC);
840                 }
841
842                 nr_reclaimed += nr_freed;
843                 local_irq_disable();
844                 if (current_is_kswapd()) {
845                         __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
846                         __count_vm_events(KSWAPD_STEAL, nr_freed);
847                 } else
848                         __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
849                 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
850
851                 if (nr_taken == 0)
852                         goto done;
853
854                 spin_lock(&zone->lru_lock);
855                 /*
856                  * Put back any unfreeable pages.
857                  */
858                 while (!list_empty(&page_list)) {
859                         page = lru_to_page(&page_list);
860                         VM_BUG_ON(PageLRU(page));
861                         SetPageLRU(page);
862                         list_del(&page->lru);
863                         if (PageActive(page))
864                                 add_page_to_active_list(zone, page);
865                         else
866                                 add_page_to_inactive_list(zone, page);
867                         if (!pagevec_add(&pvec, page)) {
868                                 spin_unlock_irq(&zone->lru_lock);
869                                 __pagevec_release(&pvec);
870                                 spin_lock_irq(&zone->lru_lock);
871                         }
872                 }
873         } while (nr_scanned < max_scan);
874         spin_unlock(&zone->lru_lock);
875 done:
876         local_irq_enable();
877         pagevec_release(&pvec);
878         return nr_reclaimed;
879 }
880
881 /*
882  * We are about to scan this zone at a certain priority level.  If that priority
883  * level is smaller (ie: more urgent) than the previous priority, then note
884  * that priority level within the zone.  This is done so that when the next
885  * process comes in to scan this zone, it will immediately start out at this
886  * priority level rather than having to build up its own scanning priority.
887  * Here, this priority affects only the reclaim-mapped threshold.
888  */
889 static inline void note_zone_scanning_priority(struct zone *zone, int priority)
890 {
891         if (priority < zone->prev_priority)
892                 zone->prev_priority = priority;
893 }
894
895 static inline int zone_is_near_oom(struct zone *zone)
896 {
897         return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
898                                 + zone_page_state(zone, NR_INACTIVE))*3;
899 }
900
901 /*
902  * This moves pages from the active list to the inactive list.
903  *
904  * We move them the other way if the page is referenced by one or more
905  * processes, from rmap.
906  *
907  * If the pages are mostly unmapped, the processing is fast and it is
908  * appropriate to hold zone->lru_lock across the whole operation.  But if
909  * the pages are mapped, the processing is slow (page_referenced()) so we
910  * should drop zone->lru_lock around each page.  It's impossible to balance
911  * this, so instead we remove the pages from the LRU while processing them.
912  * It is safe to rely on PG_active against the non-LRU pages in here because
913  * nobody will play with that bit on a non-LRU page.
914  *
915  * The downside is that we have to touch page->_count against each page.
916  * But we had to alter page->flags anyway.
917  */
918 static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
919                                 struct scan_control *sc, int priority)
920 {
921         unsigned long pgmoved;
922         int pgdeactivate = 0;
923         unsigned long pgscanned;
924         LIST_HEAD(l_hold);      /* The pages which were snipped off */
925         LIST_HEAD(l_inactive);  /* Pages to go onto the inactive_list */
926         LIST_HEAD(l_active);    /* Pages to go onto the active_list */
927         struct page *page;
928         struct pagevec pvec;
929         int reclaim_mapped = 0;
930
931         if (sc->may_swap) {
932                 long mapped_ratio;
933                 long distress;
934                 long swap_tendency;
935                 long imbalance;
936
937                 if (zone_is_near_oom(zone))
938                         goto force_reclaim_mapped;
939
940                 /*
941                  * `distress' is a measure of how much trouble we're having
942                  * reclaiming pages.  0 -> no problems.  100 -> great trouble.
943                  */
944                 distress = 100 >> min(zone->prev_priority, priority);
945
946                 /*
947                  * The point of this algorithm is to decide when to start
948                  * reclaiming mapped memory instead of just pagecache.  Work out
949                  * how much memory
950                  * is mapped.
951                  */
952                 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
953                                 global_page_state(NR_ANON_PAGES)) * 100) /
954                                         vm_total_pages;
955
956                 /*
957                  * Now decide how much we really want to unmap some pages.  The
958                  * mapped ratio is downgraded - just because there's a lot of
959                  * mapped memory doesn't necessarily mean that page reclaim
960                  * isn't succeeding.
961                  *
962                  * The distress ratio is important - we don't want to start
963                  * going oom.
964                  *
965                  * A 100% value of vm_swappiness overrides this algorithm
966                  * altogether.
967                  */
968                 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
969
970                 /*
971                  * If there's huge imbalance between active and inactive
972                  * (think active 100 times larger than inactive) we should
973                  * become more permissive, or the system will take too much
974                  * cpu before it start swapping during memory pressure.
975                  * Distress is about avoiding early-oom, this is about
976                  * making swappiness graceful despite setting it to low
977                  * values.
978                  *
979                  * Avoid div by zero with nr_inactive+1, and max resulting
980                  * value is vm_total_pages.
981                  */
982                 imbalance  = zone_page_state(zone, NR_ACTIVE);
983                 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
984
985                 /*
986                  * Reduce the effect of imbalance if swappiness is low,
987                  * this means for a swappiness very low, the imbalance
988                  * must be much higher than 100 for this logic to make
989                  * the difference.
990                  *
991                  * Max temporary value is vm_total_pages*100.
992                  */
993                 imbalance *= (vm_swappiness + 1);
994                 imbalance /= 100;
995
996                 /*
997                  * If not much of the ram is mapped, makes the imbalance
998                  * less relevant, it's high priority we refill the inactive
999                  * list with mapped pages only in presence of high ratio of
1000                  * mapped pages.
1001                  *
1002                  * Max temporary value is vm_total_pages*100.
1003                  */
1004                 imbalance *= mapped_ratio;
1005                 imbalance /= 100;
1006
1007                 /* apply imbalance feedback to swap_tendency */
1008                 swap_tendency += imbalance;
1009
1010                 /*
1011                  * Now use this metric to decide whether to start moving mapped
1012                  * memory onto the inactive list.
1013                  */
1014                 if (swap_tendency >= 100)
1015 force_reclaim_mapped:
1016                         reclaim_mapped = 1;
1017         }
1018
1019         lru_add_drain();
1020         spin_lock_irq(&zone->lru_lock);
1021         pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
1022                             &l_hold, &pgscanned, sc->order, ISOLATE_ACTIVE);
1023         zone->pages_scanned += pgscanned;
1024         __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
1025         spin_unlock_irq(&zone->lru_lock);
1026
1027         while (!list_empty(&l_hold)) {
1028                 cond_resched();
1029                 page = lru_to_page(&l_hold);
1030                 list_del(&page->lru);
1031                 if (page_mapped(page)) {
1032                         if (!reclaim_mapped ||
1033                             (total_swap_pages == 0 && PageAnon(page)) ||
1034                             page_referenced(page, 0)) {
1035                                 list_add(&page->lru, &l_active);
1036                                 continue;
1037                         }
1038                 }
1039                 list_add(&page->lru, &l_inactive);
1040         }
1041
1042         pagevec_init(&pvec, 1);
1043         pgmoved = 0;
1044         spin_lock_irq(&zone->lru_lock);
1045         while (!list_empty(&l_inactive)) {
1046                 page = lru_to_page(&l_inactive);
1047                 prefetchw_prev_lru_page(page, &l_inactive, flags);
1048                 VM_BUG_ON(PageLRU(page));
1049                 SetPageLRU(page);
1050                 VM_BUG_ON(!PageActive(page));
1051                 ClearPageActive(page);
1052
1053                 list_move(&page->lru, &zone->inactive_list);
1054                 pgmoved++;
1055                 if (!pagevec_add(&pvec, page)) {
1056                         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1057                         spin_unlock_irq(&zone->lru_lock);
1058                         pgdeactivate += pgmoved;
1059                         pgmoved = 0;
1060                         if (buffer_heads_over_limit)
1061                                 pagevec_strip(&pvec);
1062                         __pagevec_release(&pvec);
1063                         spin_lock_irq(&zone->lru_lock);
1064                 }
1065         }
1066         __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
1067         pgdeactivate += pgmoved;
1068         if (buffer_heads_over_limit) {
1069                 spin_unlock_irq(&zone->lru_lock);
1070                 pagevec_strip(&pvec);
1071                 spin_lock_irq(&zone->lru_lock);
1072         }
1073
1074         pgmoved = 0;
1075         while (!list_empty(&l_active)) {
1076                 page = lru_to_page(&l_active);
1077                 prefetchw_prev_lru_page(page, &l_active, flags);
1078                 VM_BUG_ON(PageLRU(page));
1079                 SetPageLRU(page);
1080                 VM_BUG_ON(!PageActive(page));
1081                 list_move(&page->lru, &zone->active_list);
1082                 pgmoved++;
1083                 if (!pagevec_add(&pvec, page)) {
1084                         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1085                         pgmoved = 0;
1086                         spin_unlock_irq(&zone->lru_lock);
1087                         __pagevec_release(&pvec);
1088                         spin_lock_irq(&zone->lru_lock);
1089                 }
1090         }
1091         __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
1092
1093         __count_zone_vm_events(PGREFILL, zone, pgscanned);
1094         __count_vm_events(PGDEACTIVATE, pgdeactivate);
1095         spin_unlock_irq(&zone->lru_lock);
1096
1097         pagevec_release(&pvec);
1098 }
1099
1100 /*
1101  * This is a basic per-zone page freer.  Used by both kswapd and direct reclaim.
1102  */
1103 static unsigned long shrink_zone(int priority, struct zone *zone,
1104                                 struct scan_control *sc)
1105 {
1106         unsigned long nr_active;
1107         unsigned long nr_inactive;
1108         unsigned long nr_to_scan;
1109         unsigned long nr_reclaimed = 0;
1110
1111         /*
1112          * Add one to `nr_to_scan' just to make sure that the kernel will
1113          * slowly sift through the active list.
1114          */
1115         zone->nr_scan_active +=
1116                 (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
1117         nr_active = zone->nr_scan_active;
1118         if (nr_active >= sc->swap_cluster_max)
1119                 zone->nr_scan_active = 0;
1120         else
1121                 nr_active = 0;
1122
1123         zone->nr_scan_inactive +=
1124                 (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
1125         nr_inactive = zone->nr_scan_inactive;
1126         if (nr_inactive >= sc->swap_cluster_max)
1127                 zone->nr_scan_inactive = 0;
1128         else
1129                 nr_inactive = 0;
1130
1131         while (nr_active || nr_inactive) {
1132                 if (nr_active) {
1133                         nr_to_scan = min(nr_active,
1134                                         (unsigned long)sc->swap_cluster_max);
1135                         nr_active -= nr_to_scan;
1136                         shrink_active_list(nr_to_scan, zone, sc, priority);
1137                 }
1138
1139                 if (nr_inactive) {
1140                         nr_to_scan = min(nr_inactive,
1141                                         (unsigned long)sc->swap_cluster_max);
1142                         nr_inactive -= nr_to_scan;
1143                         nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
1144                                                                 sc);
1145                 }
1146         }
1147
1148         throttle_vm_writeout(sc->gfp_mask);
1149         return nr_reclaimed;
1150 }
1151
1152 /*
1153  * This is the direct reclaim path, for page-allocating processes.  We only
1154  * try to reclaim pages from zones which will satisfy the caller's allocation
1155  * request.
1156  *
1157  * We reclaim from a zone even if that zone is over pages_high.  Because:
1158  * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1159  *    allocation or
1160  * b) The zones may be over pages_high but they must go *over* pages_high to
1161  *    satisfy the `incremental min' zone defense algorithm.
1162  *
1163  * Returns the number of reclaimed pages.
1164  *
1165  * If a zone is deemed to be full of pinned pages then just give it a light
1166  * scan then give up on it.
1167  */
1168 static unsigned long shrink_zones(int priority, struct zone **zones,
1169                                         struct scan_control *sc)
1170 {
1171         unsigned long nr_reclaimed = 0;
1172         int i;
1173
1174         sc->all_unreclaimable = 1;
1175         for (i = 0; zones[i] != NULL; i++) {
1176                 struct zone *zone = zones[i];
1177
1178                 if (!populated_zone(zone))
1179                         continue;
1180
1181                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1182                         continue;
1183
1184                 note_zone_scanning_priority(zone, priority);
1185
1186                 if (zone_is_all_unreclaimable(zone) && priority != DEF_PRIORITY)
1187                         continue;       /* Let kswapd poll it */
1188
1189                 sc->all_unreclaimable = 0;
1190
1191                 nr_reclaimed += shrink_zone(priority, zone, sc);
1192         }
1193         return nr_reclaimed;
1194 }
1195  
1196 /*
1197  * This is the main entry point to direct page reclaim.
1198  *
1199  * If a full scan of the inactive list fails to free enough memory then we
1200  * are "out of memory" and something needs to be killed.
1201  *
1202  * If the caller is !__GFP_FS then the probability of a failure is reasonably
1203  * high - the zone may be full of dirty or under-writeback pages, which this
1204  * caller can't do much about.  We kick pdflush and take explicit naps in the
1205  * hope that some of these pages can be written.  But if the allocating task
1206  * holds filesystem locks which prevent writeout this might not work, and the
1207  * allocation attempt will fail.
1208  */
1209 unsigned long try_to_free_pages(struct zone **zones, int order, gfp_t gfp_mask)
1210 {
1211         int priority;
1212         int ret = 0;
1213         unsigned long total_scanned = 0;
1214         unsigned long nr_reclaimed = 0;
1215         struct reclaim_state *reclaim_state = current->reclaim_state;
1216         unsigned long lru_pages = 0;
1217         int i;
1218         struct scan_control sc = {
1219                 .gfp_mask = gfp_mask,
1220                 .may_writepage = !laptop_mode,
1221                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1222                 .may_swap = 1,
1223                 .swappiness = vm_swappiness,
1224                 .order = order,
1225         };
1226
1227         count_vm_event(ALLOCSTALL);
1228
1229         for (i = 0; zones[i] != NULL; i++) {
1230                 struct zone *zone = zones[i];
1231
1232                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1233                         continue;
1234
1235                 lru_pages += zone_page_state(zone, NR_ACTIVE)
1236                                 + zone_page_state(zone, NR_INACTIVE);
1237         }
1238
1239         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1240                 sc.nr_scanned = 0;
1241                 if (!priority)
1242                         disable_swap_token();
1243                 nr_reclaimed += shrink_zones(priority, zones, &sc);
1244                 shrink_slab(sc.nr_scanned, gfp_mask, lru_pages);
1245                 if (reclaim_state) {
1246                         nr_reclaimed += reclaim_state->reclaimed_slab;
1247                         reclaim_state->reclaimed_slab = 0;
1248                 }
1249                 total_scanned += sc.nr_scanned;
1250                 if (nr_reclaimed >= sc.swap_cluster_max) {
1251                         ret = 1;
1252                         goto out;
1253                 }
1254
1255                 /*
1256                  * Try to write back as many pages as we just scanned.  This
1257                  * tends to cause slow streaming writers to write data to the
1258                  * disk smoothly, at the dirtying rate, which is nice.   But
1259                  * that's undesirable in laptop mode, where we *want* lumpy
1260                  * writeout.  So in laptop mode, write out the whole world.
1261                  */
1262                 if (total_scanned > sc.swap_cluster_max +
1263                                         sc.swap_cluster_max / 2) {
1264                         wakeup_pdflush(laptop_mode ? 0 : total_scanned);
1265                         sc.may_writepage = 1;
1266                 }
1267
1268                 /* Take a nap, wait for some writeback to complete */
1269                 if (sc.nr_scanned && priority < DEF_PRIORITY - 2)
1270                         congestion_wait(WRITE, HZ/10);
1271         }
1272         /* top priority shrink_caches still had more to do? don't OOM, then */
1273         if (!sc.all_unreclaimable)
1274                 ret = 1;
1275 out:
1276         /*
1277          * Now that we've scanned all the zones at this priority level, note
1278          * that level within the zone so that the next thread which performs
1279          * scanning of this zone will immediately start out at this priority
1280          * level.  This affects only the decision whether or not to bring
1281          * mapped pages onto the inactive list.
1282          */
1283         if (priority < 0)
1284                 priority = 0;
1285         for (i = 0; zones[i] != NULL; i++) {
1286                 struct zone *zone = zones[i];
1287
1288                 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1289                         continue;
1290
1291                 zone->prev_priority = priority;
1292         }
1293         return ret;
1294 }
1295
1296 /*
1297  * For kswapd, balance_pgdat() will work across all this node's zones until
1298  * they are all at pages_high.
1299  *
1300  * Returns the number of pages which were actually freed.
1301  *
1302  * There is special handling here for zones which are full of pinned pages.
1303  * This can happen if the pages are all mlocked, or if they are all used by
1304  * device drivers (say, ZONE_DMA).  Or if they are all in use by hugetlb.
1305  * What we do is to detect the case where all pages in the zone have been
1306  * scanned twice and there has been zero successful reclaim.  Mark the zone as
1307  * dead and from now on, only perform a short scan.  Basically we're polling
1308  * the zone for when the problem goes away.
1309  *
1310  * kswapd scans the zones in the highmem->normal->dma direction.  It skips
1311  * zones which have free_pages > pages_high, but once a zone is found to have
1312  * free_pages <= pages_high, we scan that zone and the lower zones regardless
1313  * of the number of free pages in the lower zones.  This interoperates with
1314  * the page allocator fallback scheme to ensure that aging of pages is balanced
1315  * across the zones.
1316  */
1317 static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
1318 {
1319         int all_zones_ok;
1320         int priority;
1321         int i;
1322         unsigned long total_scanned;
1323         unsigned long nr_reclaimed;
1324         struct reclaim_state *reclaim_state = current->reclaim_state;
1325         struct scan_control sc = {
1326                 .gfp_mask = GFP_KERNEL,
1327                 .may_swap = 1,
1328                 .swap_cluster_max = SWAP_CLUSTER_MAX,
1329                 .swappiness = vm_swappiness,
1330                 .order = order,
1331         };
1332         /*
1333          * temp_priority is used to remember the scanning priority at which
1334          * this zone was successfully refilled to free_pages == pages_high.
1335          */
1336         int temp_priority[MAX_NR_ZONES];
1337
1338 loop_again:
1339         total_scanned = 0;
1340         nr_reclaimed = 0;
1341         sc.may_writepage = !laptop_mode;
1342         count_vm_event(PAGEOUTRUN);
1343
1344         for (i = 0; i < pgdat->nr_zones; i++)
1345                 temp_priority[i] = DEF_PRIORITY;
1346
1347         for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1348                 int end_zone = 0;       /* Inclusive.  0 = ZONE_DMA */
1349                 unsigned long lru_pages = 0;
1350
1351                 /* The swap token gets in the way of swapout... */
1352                 if (!priority)
1353                         disable_swap_token();
1354
1355                 all_zones_ok = 1;
1356
1357                 /*
1358                  * Scan in the highmem->dma direction for the highest
1359                  * zone which needs scanning
1360                  */
1361                 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1362                         struct zone *zone = pgdat->node_zones + i;
1363
1364                         if (!populated_zone(zone))
1365                                 continue;
1366
1367                         if (zone_is_all_unreclaimable(zone) &&
1368                             priority != DEF_PRIORITY)
1369                                 continue;
1370
1371                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1372                                                0, 0)) {
1373                                 end_zone = i;
1374                                 break;
1375                         }
1376                 }
1377                 if (i < 0)
1378                         goto out;
1379
1380                 for (i = 0; i <= end_zone; i++) {
1381                         struct zone *zone = pgdat->node_zones + i;
1382
1383                         lru_pages += zone_page_state(zone, NR_ACTIVE)
1384                                         + zone_page_state(zone, NR_INACTIVE);
1385                 }
1386
1387                 /*
1388                  * Now scan the zone in the dma->highmem direction, stopping
1389                  * at the last zone which needs scanning.
1390                  *
1391                  * We do this because the page allocator works in the opposite
1392                  * direction.  This prevents the page allocator from allocating
1393                  * pages behind kswapd's direction of progress, which would
1394                  * cause too much scanning of the lower zones.
1395                  */
1396                 for (i = 0; i <= end_zone; i++) {
1397                         struct zone *zone = pgdat->node_zones + i;
1398                         int nr_slab;
1399
1400                         if (!populated_zone(zone))
1401                                 continue;
1402
1403                         if (zone_is_all_unreclaimable(zone) &&
1404                                         priority != DEF_PRIORITY)
1405                                 continue;
1406
1407                         if (!zone_watermark_ok(zone, order, zone->pages_high,
1408                                                end_zone, 0))
1409                                 all_zones_ok = 0;
1410                         temp_priority[i] = priority;
1411                         sc.nr_scanned = 0;
1412                         note_zone_scanning_priority(zone, priority);
1413                         /*
1414                          * We put equal pressure on every zone, unless one
1415                          * zone has way too many pages free already.
1416                          */
1417                         if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1418                                                 end_zone, 0))
1419                                 nr_reclaimed += shrink_zone(priority, zone, &sc);
1420                         reclaim_state->reclaimed_slab = 0;
1421                         nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1422                                                 lru_pages);
1423                         nr_reclaimed += reclaim_state->reclaimed_slab;
1424                         total_scanned += sc.nr_scanned;
1425                         if (zone_is_all_unreclaimable(zone))
1426                                 continue;
1427                         if (nr_slab == 0 && zone->pages_scanned >=
1428                                 (zone_page_state(zone, NR_ACTIVE)
1429                                 + zone_page_state(zone, NR_INACTIVE)) * 6)
1430                                         zone_set_flag(zone,
1431                                                       ZONE_ALL_UNRECLAIMABLE);
1432                         /*
1433                          * If we've done a decent amount of scanning and
1434                          * the reclaim ratio is low, start doing writepage
1435                          * even in laptop mode
1436                          */
1437                         if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
1438                             total_scanned > nr_reclaimed + nr_reclaimed / 2)
1439                                 sc.may_writepage = 1;
1440                 }
1441                 if (all_zones_ok)
1442                         break;          /* kswapd: all done */
1443                 /*
1444                  * OK, kswapd is getting into trouble.  Take a nap, then take
1445                  * another pass across the zones.
1446                  */
1447                 if (total_scanned && priority < DEF_PRIORITY - 2)
1448                         congestion_wait(WRITE, HZ/10);
1449
1450                 /*
1451                  * We do this so kswapd doesn't build up large priorities for
1452                  * example when it is freeing in parallel with allocators. It
1453                  * matches the direct reclaim path behaviour in terms of impact
1454                  * on zone->*_priority.
1455                  */
1456                 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
1457                         break;
1458         }
1459 out:
1460         /*
1461          * Note within each zone the priority level at which this zone was
1462          * brought into a happy state.  So that the next thread which scans this
1463          * zone will start out at that priority level.
1464          */
1465         for (i = 0; i < pgdat->nr_zones; i++) {
1466                 struct zone *zone = pgdat->node_zones + i;
1467
1468                 zone->prev_priority = temp_priority[i];
1469         }
1470         if (!all_zones_ok) {
1471                 cond_resched();
1472
1473                 try_to_freeze();
1474
1475                 goto loop_again;
1476         }
1477
1478         return nr_reclaimed;
1479 }
1480
1481 /*
1482  * The background pageout daemon, started as a kernel thread
1483  * from the init process. 
1484  *
1485  * This basically trickles out pages so that we have _some_
1486  * free memory available even if there is no other activity
1487  * that frees anything up. This is needed for things like routing
1488  * etc, where we otherwise might have all activity going on in
1489  * asynchronous contexts that cannot page things out.
1490  *
1491  * If there are applications that are active memory-allocators
1492  * (most normal use), this basically shouldn't matter.
1493  */
1494 static int kswapd(void *p)
1495 {
1496         unsigned long order;
1497         pg_data_t *pgdat = (pg_data_t*)p;
1498         struct task_struct *tsk = current;
1499         DEFINE_WAIT(wait);
1500         struct reclaim_state reclaim_state = {
1501                 .reclaimed_slab = 0,
1502         };
1503         cpumask_t cpumask;
1504
1505         cpumask = node_to_cpumask(pgdat->node_id);
1506         if (!cpus_empty(cpumask))
1507                 set_cpus_allowed(tsk, cpumask);
1508         current->reclaim_state = &reclaim_state;
1509
1510         /*
1511          * Tell the memory management that we're a "memory allocator",
1512          * and that if we need more memory we should get access to it
1513          * regardless (see "__alloc_pages()"). "kswapd" should
1514          * never get caught in the normal page freeing logic.
1515          *
1516          * (Kswapd normally doesn't need memory anyway, but sometimes
1517          * you need a small amount of memory in order to be able to
1518          * page out something else, and this flag essentially protects
1519          * us from recursively trying to free more memory as we're
1520          * trying to free the first piece of memory in the first place).
1521          */
1522         tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
1523         set_freezable();
1524
1525         order = 0;
1526         for ( ; ; ) {
1527                 unsigned long new_order;
1528
1529                 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1530                 new_order = pgdat->kswapd_max_order;
1531                 pgdat->kswapd_max_order = 0;
1532                 if (order < new_order) {
1533                         /*
1534                          * Don't sleep if someone wants a larger 'order'
1535                          * allocation
1536                          */
1537                         order = new_order;
1538                 } else {
1539                         if (!freezing(current))
1540                                 schedule();
1541
1542                         order = pgdat->kswapd_max_order;
1543                 }
1544                 finish_wait(&pgdat->kswapd_wait, &wait);
1545
1546                 if (!try_to_freeze()) {
1547                         /* We can speed up thawing tasks if we don't call
1548                          * balance_pgdat after returning from the refrigerator
1549                          */
1550                         balance_pgdat(pgdat, order);
1551                 }
1552         }
1553         return 0;
1554 }
1555
1556 /*
1557  * A zone is low on free memory, so wake its kswapd task to service it.
1558  */
1559 void wakeup_kswapd(struct zone *zone, int order)
1560 {
1561         pg_data_t *pgdat;
1562
1563         if (!populated_zone(zone))
1564                 return;
1565
1566         pgdat = zone->zone_pgdat;
1567         if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
1568                 return;
1569         if (pgdat->kswapd_max_order < order)
1570                 pgdat->kswapd_max_order = order;
1571         if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1572                 return;
1573         if (!waitqueue_active(&pgdat->kswapd_wait))
1574                 return;
1575         wake_up_interruptible(&pgdat->kswapd_wait);
1576 }
1577
1578 #ifdef CONFIG_PM
1579 /*
1580  * Helper function for shrink_all_memory().  Tries to reclaim 'nr_pages' pages
1581  * from LRU lists system-wide, for given pass and priority, and returns the
1582  * number of reclaimed pages
1583  *
1584  * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1585  */
1586 static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1587                                       int pass, struct scan_control *sc)
1588 {
1589         struct zone *zone;
1590         unsigned long nr_to_scan, ret = 0;
1591
1592         for_each_zone(zone) {
1593
1594                 if (!populated_zone(zone))
1595                         continue;
1596
1597                 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
1598                         continue;
1599
1600                 /* For pass = 0 we don't shrink the active list */
1601                 if (pass > 0) {
1602                         zone->nr_scan_active +=
1603                                 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
1604                         if (zone->nr_scan_active >= nr_pages || pass > 3) {
1605                                 zone->nr_scan_active = 0;
1606                                 nr_to_scan = min(nr_pages,
1607                                         zone_page_state(zone, NR_ACTIVE));
1608                                 shrink_active_list(nr_to_scan, zone, sc, prio);
1609                         }
1610                 }
1611
1612                 zone->nr_scan_inactive +=
1613                         (zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
1614                 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1615                         zone->nr_scan_inactive = 0;
1616                         nr_to_scan = min(nr_pages,
1617                                 zone_page_state(zone, NR_INACTIVE));
1618                         ret += shrink_inactive_list(nr_to_scan, zone, sc);
1619                         if (ret >= nr_pages)
1620                                 return ret;
1621                 }
1622         }
1623
1624         return ret;
1625 }
1626
1627 static unsigned long count_lru_pages(void)
1628 {
1629         return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
1630 }
1631
1632 /*
1633  * Try to free `nr_pages' of memory, system-wide, and return the number of
1634  * freed pages.
1635  *
1636  * Rather than trying to age LRUs the aim is to preserve the overall
1637  * LRU order by reclaiming preferentially
1638  * inactive > active > active referenced > active mapped
1639  */
1640 unsigned long shrink_all_memory(unsigned long nr_pages)
1641 {
1642         unsigned long lru_pages, nr_slab;
1643         unsigned long ret = 0;
1644         int pass;
1645         struct reclaim_state reclaim_state;
1646         struct scan_control sc = {
1647                 .gfp_mask = GFP_KERNEL,
1648                 .may_swap = 0,
1649                 .swap_cluster_max = nr_pages,
1650                 .may_writepage = 1,
1651                 .swappiness = vm_swappiness,
1652         };
1653
1654         current->reclaim_state = &reclaim_state;
1655
1656         lru_pages = count_lru_pages();
1657         nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
1658         /* If slab caches are huge, it's better to hit them first */
1659         while (nr_slab >= lru_pages) {
1660                 reclaim_state.reclaimed_slab = 0;
1661                 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1662                 if (!reclaim_state.reclaimed_slab)
1663                         break;
1664
1665                 ret += reclaim_state.reclaimed_slab;
1666                 if (ret >= nr_pages)
1667                         goto out;
1668
1669                 nr_slab -= reclaim_state.reclaimed_slab;
1670         }
1671
1672         /*
1673          * We try to shrink LRUs in 5 passes:
1674          * 0 = Reclaim from inactive_list only
1675          * 1 = Reclaim from active list but don't reclaim mapped
1676          * 2 = 2nd pass of type 1
1677          * 3 = Reclaim mapped (normal reclaim)
1678          * 4 = 2nd pass of type 3
1679          */
1680         for (pass = 0; pass < 5; pass++) {
1681                 int prio;
1682
1683                 /* Force reclaiming mapped pages in the passes #3 and #4 */
1684                 if (pass > 2) {
1685                         sc.may_swap = 1;
1686                         sc.swappiness = 100;
1687                 }
1688
1689                 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1690                         unsigned long nr_to_scan = nr_pages - ret;
1691
1692                         sc.nr_scanned = 0;
1693                         ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1694                         if (ret >= nr_pages)
1695                                 goto out;
1696
1697                         reclaim_state.reclaimed_slab = 0;
1698                         shrink_slab(sc.nr_scanned, sc.gfp_mask,
1699                                         count_lru_pages());
1700                         ret += reclaim_state.reclaimed_slab;
1701                         if (ret >= nr_pages)
1702                                 goto out;
1703
1704                         if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
1705                                 congestion_wait(WRITE, HZ / 10);
1706                 }
1707         }
1708
1709         /*
1710          * If ret = 0, we could not shrink LRUs, but there may be something
1711          * in slab caches
1712          */
1713         if (!ret) {
1714                 do {
1715                         reclaim_state.reclaimed_slab = 0;
1716                         shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
1717                         ret += reclaim_state.reclaimed_slab;
1718                 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
1719         }
1720
1721 out:
1722         current->reclaim_state = NULL;
1723
1724         return ret;
1725 }
1726 #endif
1727
1728 /* It's optimal to keep kswapds on the same CPUs as their memory, but
1729    not required for correctness.  So if the last cpu in a node goes
1730    away, we get changed to run anywhere: as the first one comes back,
1731    restore their cpu bindings. */
1732 static int __devinit cpu_callback(struct notifier_block *nfb,
1733                                   unsigned long action, void *hcpu)
1734 {
1735         pg_data_t *pgdat;
1736         cpumask_t mask;
1737         int nid;
1738
1739         if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
1740                 for_each_node_state(nid, N_HIGH_MEMORY) {
1741                         pgdat = NODE_DATA(nid);
1742                         mask = node_to_cpumask(pgdat->node_id);
1743                         if (any_online_cpu(mask) != NR_CPUS)
1744                                 /* One of our CPUs online: restore mask */
1745                                 set_cpus_allowed(pgdat->kswapd, mask);
1746                 }
1747         }
1748         return NOTIFY_OK;
1749 }
1750
1751 /*
1752  * This kswapd start function will be called by init and node-hot-add.
1753  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1754  */
1755 int kswapd_run(int nid)
1756 {
1757         pg_data_t *pgdat = NODE_DATA(nid);
1758         int ret = 0;
1759
1760         if (pgdat->kswapd)
1761                 return 0;
1762
1763         pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1764         if (IS_ERR(pgdat->kswapd)) {
1765                 /* failure at boot is fatal */
1766                 BUG_ON(system_state == SYSTEM_BOOTING);
1767                 printk("Failed to start kswapd on node %d\n",nid);
1768                 ret = -1;
1769         }
1770         return ret;
1771 }
1772
1773 static int __init kswapd_init(void)
1774 {
1775         int nid;
1776
1777         swap_setup();
1778         for_each_node_state(nid, N_HIGH_MEMORY)
1779                 kswapd_run(nid);
1780         hotcpu_notifier(cpu_callback, 0);
1781         return 0;
1782 }
1783
1784 module_init(kswapd_init)
1785
1786 #ifdef CONFIG_NUMA
1787 /*
1788  * Zone reclaim mode
1789  *
1790  * If non-zero call zone_reclaim when the number of free pages falls below
1791  * the watermarks.
1792  */
1793 int zone_reclaim_mode __read_mostly;
1794
1795 #define RECLAIM_OFF 0
1796 #define RECLAIM_ZONE (1<<0)     /* Run shrink_cache on the zone */
1797 #define RECLAIM_WRITE (1<<1)    /* Writeout pages during reclaim */
1798 #define RECLAIM_SWAP (1<<2)     /* Swap pages out during reclaim */
1799
1800 /*
1801  * Priority for ZONE_RECLAIM. This determines the fraction of pages
1802  * of a node considered for each zone_reclaim. 4 scans 1/16th of
1803  * a zone.
1804  */
1805 #define ZONE_RECLAIM_PRIORITY 4
1806
1807 /*
1808  * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1809  * occur.
1810  */
1811 int sysctl_min_unmapped_ratio = 1;
1812
1813 /*
1814  * If the number of slab pages in a zone grows beyond this percentage then
1815  * slab reclaim needs to occur.
1816  */
1817 int sysctl_min_slab_ratio = 5;
1818
1819 /*
1820  * Try to free up some pages from this zone through reclaim.
1821  */
1822 static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1823 {
1824         /* Minimum pages needed in order to stay on node */
1825         const unsigned long nr_pages = 1 << order;
1826         struct task_struct *p = current;
1827         struct reclaim_state reclaim_state;
1828         int priority;
1829         unsigned long nr_reclaimed = 0;
1830         struct scan_control sc = {
1831                 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1832                 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
1833                 .swap_cluster_max = max_t(unsigned long, nr_pages,
1834                                         SWAP_CLUSTER_MAX),
1835                 .gfp_mask = gfp_mask,
1836                 .swappiness = vm_swappiness,
1837         };
1838         unsigned long slab_reclaimable;
1839
1840         disable_swap_token();
1841         cond_resched();
1842         /*
1843          * We need to be able to allocate from the reserves for RECLAIM_SWAP
1844          * and we also need to be able to write out pages for RECLAIM_WRITE
1845          * and RECLAIM_SWAP.
1846          */
1847         p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
1848         reclaim_state.reclaimed_slab = 0;
1849         p->reclaim_state = &reclaim_state;
1850
1851         if (zone_page_state(zone, NR_FILE_PAGES) -
1852                 zone_page_state(zone, NR_FILE_MAPPED) >
1853                 zone->min_unmapped_pages) {
1854                 /*
1855                  * Free memory by calling shrink zone with increasing
1856                  * priorities until we have enough memory freed.
1857                  */
1858                 priority = ZONE_RECLAIM_PRIORITY;
1859                 do {
1860                         note_zone_scanning_priority(zone, priority);
1861                         nr_reclaimed += shrink_zone(priority, zone, &sc);
1862                         priority--;
1863                 } while (priority >= 0 && nr_reclaimed < nr_pages);
1864         }
1865
1866         slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1867         if (slab_reclaimable > zone->min_slab_pages) {
1868                 /*
1869                  * shrink_slab() does not currently allow us to determine how
1870                  * many pages were freed in this zone. So we take the current
1871                  * number of slab pages and shake the slab until it is reduced
1872                  * by the same nr_pages that we used for reclaiming unmapped
1873                  * pages.
1874                  *
1875                  * Note that shrink_slab will free memory on all zones and may
1876                  * take a long time.
1877                  */
1878                 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
1879                         zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
1880                                 slab_reclaimable - nr_pages)
1881                         ;
1882
1883                 /*
1884                  * Update nr_reclaimed by the number of slab pages we
1885                  * reclaimed from this zone.
1886                  */
1887                 nr_reclaimed += slab_reclaimable -
1888                         zone_page_state(zone, NR_SLAB_RECLAIMABLE);
1889         }
1890
1891         p->reclaim_state = NULL;
1892         current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
1893         return nr_reclaimed >= nr_pages;
1894 }
1895
1896 int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
1897 {
1898         int node_id;
1899         int ret;
1900
1901         /*
1902          * Zone reclaim reclaims unmapped file backed pages and
1903          * slab pages if we are over the defined limits.
1904          *
1905          * A small portion of unmapped file backed pages is needed for
1906          * file I/O otherwise pages read by file I/O will be immediately
1907          * thrown out if the zone is overallocated. So we do not reclaim
1908          * if less than a specified percentage of the zone is used by
1909          * unmapped file backed pages.
1910          */
1911         if (zone_page_state(zone, NR_FILE_PAGES) -
1912             zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
1913             && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
1914                         <= zone->min_slab_pages)
1915                 return 0;
1916
1917         if (zone_is_all_unreclaimable(zone))
1918                 return 0;
1919
1920         /*
1921          * Do not scan if the allocation should not be delayed.
1922          */
1923         if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
1924                         return 0;
1925
1926         /*
1927          * Only run zone reclaim on the local zone or on zones that do not
1928          * have associated processors. This will favor the local processor
1929          * over remote processors and spread off node memory allocations
1930          * as wide as possible.
1931          */
1932         node_id = zone_to_nid(zone);
1933         if (node_state(node_id, N_CPU) && node_id != numa_node_id())
1934                 return 0;
1935
1936         if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
1937                 return 0;
1938         ret = __zone_reclaim(zone, gfp_mask, order);
1939         zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
1940
1941         return ret;
1942 }
1943 #endif